27 Nov 2018
6 things AI can do, and 5 things it can’t (yet)
Artificial intelligence; it’s one of the tech world’s most pervasive buzzwords, and has been for quite some time.
From tech billionaires to scientists and engineers, it seems as though everyone has something different to say about AI’s amazing capabilities and its current limitations. In a world filled to bursting with AI punditry and hysterical fearmongering, separating AI fact from AI fiction can be a tough task indeed.
That’s why we here at Phrasee decided to take a look at the world’s current AI landscape and find out exactly what artificial intelligence technology is capable of at this very moment, as well as what it hasn’t quite gotten the hang of yet.
Here’s what we found…
Things AI CAN do:
1) Drive you around in relative safety
Autonomous vehicles are already being tested at scale on public roads and in several major cities. Although there have been some incidents (including 4 deaths) which have sparked debate about their efficacy, self-driving vehicles have already logged millions of miles with far lower rates of collisions and injuries than those of human drivers. While such debates are likely to follow autonomous vehicles well into the next decade, the fact remains that the progress such technologies have made over the past few years has been nothing short of astonishing.
2) Book things for you by phone
Google made waves (and headlines) this year when it demoed its natural language generation platform “Duplex” for an astonished crowd at I/O 2018. The demonstration offered a stunning sneak peek into what AI assistants like Google Home, Alexa, and Siri will soon be capable of, with recordings of Duplex booking a hair appointment and a restaurant reservation, overcoming obstacles like unexpected questions and thick accents with barely a second’s hesitation.
3) Write marketing language which outperforms that written by humans
Artistic pursuits like writing have long been thought to be beyond the reach of artificial intelligence. Here at Phrasee, we haven’t let that stand in our way. Artificial intelligence has been generating and optimizing marketing copy that consistently outperforms marketing copy written by human copywriters for more than 2 years. Although an AI-generated novel is still many moons away, our AI algorithms have proven that writing and split-testing marketing copy for subject lines, push notifications, and social media ads in a multitude of languages is a problem well within AI’s current grasp.
4) Understand you (most of the time)
As of this writing, more than 18 million homes in the United States alone now contain a smart speaker. That number is expected to grow considerably in the months and years to come. Devices like Google Home and Alexa are already redefining the human-operating system interface. Although far from perfect at this stage, such devices (when used correctly) represent a watershed moment for natural language processing technology. It is only a matter of time before they learn to understand us perfectly.
5) Predict which movies/TV shows/songs/products you’ll like with a reasonable degree of accuracy
Whenever you fire up Netflix, log on to your Spotify or glance at the “you might also like” section on an eCommerce site, you’re shaking hands with artificial intelligence. Machine learning recommendation engines are one of the most widely-used examples of AI in use today. As these engines collect data on your habits, they become increasingly effective in predicting which recommendations you will respond positively to.
6) Diagnose skin cancer more effectively than dermatologists
AI scientists have barely scratched the surface of the impressive applications AI could have in store for the medical field. A study published earlier this year in the Annals of Oncology pitted a skin cancer-diagnosing AI tool against 58 dermatologists from 17 countries to accurately differentiate cancerous skin lesions from benign skin lesions. While the dermatologists in the study identified the cancerous lesions correctly in 86.6% of cases, the cancer-detecting AI scored 95%, outperforming its human competition in almost 1 in 10 cases.
Things AI CAN’T do (yet):
1) Run a Twitter account unsupervised
Designed to mimic the linguistic patterns of a 19-year-old woman, Microsoft’s Twitter chatbot “Tay” lasted a grand total of under 16 hours before it had to be shut down by her design team. Why? Because the AI algorithm was designed to learn how to interact with humans based on the tweets it saw online, which was not an ideal plan. Tay was tweeting racist and misogynistic diatribes within hours once twitters trolls got involved.
2) Write an episode of Friends
While AI has proven adept indeed at generating amazing optimized short-form marketing copy, it hasn’t always excelled at writing longform creative. Such was the case when software developer Andy Herd unveiled a recurrent neural network he had designed tasked with generating scripts for new episodes of the hit TV show, Friends. The scripts it churned out can only really be described as gibberish, although they do make for interesting reading.
3) Gain admission into The University of Tokyo
Despite having a full 5 years to train an artificial intelligence dubbed “Todai Robot”, a team of Japanese researchers were unable to develop an AI algorithm capable of achieving scores on Japan’s unified admission test for national universities which would qualify it for entry into Japan’s top university. Despite showing steady academic improvement over its first 4 years, Todai’s academic performance between year 4 and year 5 showed almost no improvement. The team has since abandoned the project.
4) Say “no” to children
When a 6-year-old Dallas Texas girl named Brooke Neitzel spoke with her family’s Alexa Dot about her love of dollhouses and sugar cookies, it did the kindest thing it was capable of. It ordered them for her. 160 US dollars later, Brooke found herself in possession of a huge tub of cookies and one really swanky dollhouse. Her parents were unimpressed, and have since added a 4-digit code to their Alexa dot before it can make any purchases.
5) Compose a good song
While artificial intelligence is most certainly having a significant impact on the way popular music is made, human input is still very much required at this stage. Sure, there have been a few borderline listenable songs generated entirely by artificial intelligence, but songs that can actually be considered “good” remain out of AI’s reach, at least for the moment.
Although the current artificial landscape remains dotted with anecdotes about its various successes and failures, there remains little question that the technology is progressing at an incredible rate. Well-designed, task-specific artificial intelligence tools are popping up every day, and most are able to complete the tasks they’ve been assigned very, very effectively.